Deep Reinforcement Learning Control of Wave Energy Converters

Abstract

One major challenge of converting wave energy to useful electricity is the development of highly efficient Power Take-off (PTO) control algorithms. Conventional model-based controls are typically developed based on reduced-order models and neglecting the dynamics of other subsystems. Although showing promising performance in idealized conditions, it may be misleading in practice. On the other hand, it is nearly impossible to derive a model-based control for a highly nonlinear/complex system (e.g., a wave-to-wire model). Therefore, in this study, we propose a Deep Reinforcement Learning (DRL, model-free) control that aims at/capable of optimizing the performance of WECs from wave to wire. This wave-to-wire model is composed of a heaving point absorber and a direct-drive PTO unit. The numerical simulations are first conducted on comparing the performance of the proposed DRL and conventional model-based controls. The results show maximumly a 152% improvement in terms of the electricity generation and an 84% improvement in terms of the power quality. Moreover, the robustness of the proposed control is also validated under varied real ocean conditions at PacWave. The results indicate a consistent improvement of power production and quality of the proposed DRL control compared to model-based controls.